3D environmental modeling and drivable road identification for a long-range rover

摘要

It is necessary for a long-range space rover to model its 3D environment in order to identify drivable roads while avoiding obstacles and hazards. We present a method of modeling a 3D environment, while identifying drivable roads and obstacles, by accumulating a sequence of point clouds from stereo. The technical issues we deal with are how to avoid error accumulation in registering point clouds using visual odometry and how to keep the quality of the model under the constraint of limited number of points to carry. In this paper, we propose: 1) Digital Elevation Map (DEM) based global localization. 2) Level of Detail (LOD) based accumulation of local 3D point clouds. The application of our proposed method to Devon Island dataset [12] indicates that the error accumulation is contained within 5%.